goosedump

goosedump browses, searches, converts, and compacts saved coding-agent conversations. It supports Claude Code, Codex, Crush, Gemini CLI, Goose, OpenCode, and Pi.

It builds on ideas from lllyasviel/VCC.

Install

Prebuilt binaries are available for Linux x64 and arm64, macOS arm64, and Windows x64:

npm install --global @jarkkojs/goosedump

To build from source:

cargo build --release

Source builds require a C compiler for the bundled SQLite library.

Quick start

goosedump list
goosedump show pi:context-id
goosedump grep pi:context-id '*database*'
goosedump search pi:context-id 'migration failure'
goosedump compact pi:context-id

Use goosedump --help or the manual page for all commands and options.

Features

• Lists and filters contexts across all supported providers.
• Renders provider-native output and converts conversations between provider formats.
• Searches session history with BM25 ranking and glob matching.
• Imports contexts into another provider's store; pair with delete to move them.
• Recalls durable memory harvested from compacted context ranges.

Persistent memory

compact emits a summary without retaining the selected range by default. Pass --train to harvest it into the platform state directory (~/.local/state/goosedump/goosedump.db on Linux). Stored entries are classified as decisions, facts, preferences, procedures, or episodes and include trust, confidence, validity, scope, derivation, and concrete source provenance. Admission screening quarantines instruction-like, generated, and malformed content before retrieval. Indexing uses a durable pending/indexed/failed ledger, recovers interrupted work, and stops automatic retries after three attempts. Use recall --type TYPE to restrict a search, stats to inspect type, governance, and ingestion counts, govern review to inspect quarantined entries, and ingest review to inspect indexing failures.

Memory keeps historical provenance. Deleting or moving a source context does not remove previously harvested content. forget HASH and forget PROVIDER:ID remove memory and create durable tombstones that prevent re-ingestion. Use govern approve, quarantine, supersede, contradict, and expire for explicit lifecycle control. Pass compact --train only when the selected range may be retained; screening is defensive and is not a substitute for secret filtering. Reset one terminal failure for another bounded retry cycle with ingest retry HASH.

mutate performs an optional local merge of two archived memories. It normally waits for sufficient usage evidence; mutate --force is intended for testing. The fixed GPT-OSS-20B MXFP4 text model is downloaded on first use, or ahead of time with goosedump pull.

Inference performance

The custom GPT-OSS-20B CPU engine defaults to the physical CPU core count, capped by the CPUs available to the process. Override the compute thread count:

GOOSEDUMP_INFERENCE_THREADS=6 \
goosedump compact pi:context-id

GPT-OSS-20B text generation uses the custom CPU engine. Measure a representative context before changing the defaults. The bench feature builds a standalone harness that shells out to goosedump compact with GOOSEDUMP_PROFILE_COMPACT=1 and reports end-to-end and per-stage timings:

# Edit benches/bench.txt to point at a real context, then:
cargo bench --features bench

Each case emits JSON with per-run timing (including model loading), median and p95 latency, and per-stage medians and p95s. Keep the machine, context range, and model cache unchanged when comparing configurations.

Pi extension

The bundled pi-goosedump extension adds session-history tools, durable-memory recall, and goosedump compaction to Pi:

pi install npm:pi-goosedump

Development

cargo test --all-targets
cargo clippy --all-targets -- -D warnings
cargo fmt --check

The published Linux x64 and arm64 binaries are static glibc PIEs.

License

The Rust program is licensed under LGPL-2.1-or-later. The npm distribution also contains Apache-2.0-licensed components.